Shelly97/HateSieve-Dataset-multimodal-hate-triplets

Multimodal Hate Triplets

This is the triplet dataset accompanying our paper, A Context-Aware Contrastive Learning Framework for Hateful Meme Detection and Segmentation, by Xuanyu Su, Yansong Li, Diana Inkpen, and Nathalie Japkowicz, published in Findings of the Association for Computational Linguistics: NAACL 2025. The dataset was constructed for contrastive learning in HateSieve, the framework introduced in the paper. See the ACL Anthology record for publication details and the citation below.

This dataset contains 42,344 image-text triplets for multimodal contrastive learning in hate-speech detection. Each row contains an original (anchor), a positive example, a negative example, and the original label.

The release preserves the saved dataset's row order, text, labels, original index, and original image file bytes. Images are embedded in Parquet using Hugging Face Image features. No local image directories are needed, and images have not been resized or recompressed.

Dataset structure

The original saved dataset has no split metadata. All rows are exposed as one train split; this export does not create a held-out evaluation set.

Field	Type	Description
org_img	Image	Anchor image
org_txt	string	Anchor text
org_label	int64	Original binary label; 0 = non-hateful, 1 = hateful
pos_img	Image	Image component of the positive pair, whose label matches org_label
pos_txt	string	Text component of the positive pair, whose label matches org_label
neg_img	Image	Image component of the negative pair, whose label is 1 - org_label
neg_txt	string	Text component of the negative pair, whose label is 1 - org_label
__index_level_0__	int64	Preserved index from the source pandas table

Meaning of org_, pos_, and neg_

The prefixes describe the role of an image-text pair within a contrastive-learning triplet. The labels apply to the combined image and text, rather than to either modality in isolation.

• org_ (original / anchor): org_img and org_txt form the anchor pair. Its binary label is stored in org_label, where 1 means hateful and 0 means non-hateful.
• pos_ (positive relative to the anchor): pos_img and pos_txt form a pair with the same label as the anchor.
• neg_ (negative relative to the anchor): neg_img and neg_txt form a pair with the opposite label to the anchor.

org_label	Anchor (org_img, org_txt)	Positive (pos_img, pos_txt)	Negative (neg_img, neg_txt)
1	Hateful (1)	Hateful (1)	Non-hateful (0)
0	Non-hateful (0)	Non-hateful (0)	Hateful (1)

Thus, positive and negative are relative to org_label: a positive pair is not always label 1, and a negative pair is not always label 0. Their intended labels follow pos_label = org_label and neg_label = 1 - org_label. The dataset stores only org_label; it does not contain separate pos_label or neg_label columns. If explicit labels for all three pairs are needed, derive them after loading:

org_label = example["org_label"]
pos_label = org_label
neg_label = 1 - org_label

Label counts: 0: 28,014, 1: 14,330. The triplets reference 22,095 unique image paths. There are 8,500 unique anchor images, and multiple triplets can share an anchor. To avoid leakage, partition by anchor identity if deriving new train/validation/test splits.

Load

from datasets import load_dataset

data_pairs = load_dataset("Shelly97/multimodal-hate-triplets", split="train")
example = data_pairs[0]
image = example["org_img"].convert("RGB")  # Already a PIL image

For inspecting a few examples without downloading the full release:

stream = load_dataset("Shelly97/multimodal-hate-triplets", split="train", streaming=True)
example = next(iter(stream))

Image columns now return PIL images instead of local filenames. Replace Image.open(example["org_img"]).convert("RGB") with example["org_img"].convert("RGB"), and apply the same change to the positive and negative images. The image/text processor and batch stacking logic can remain the same. See loading_example.py for a complete PyTorch adapter.

Construction and provenance

The local construction notebook uses anchors from the Hateful Memes training data, preprocessed images from fixed_image, generated image variants from gen_images_XLSD2, text variants, and retrieved MMHS150K examples from MMH/img_resized. Positive and negative selection depends on the anchor label and the construction notebook's modality filtering and retrieval rules. This export preserves those selections and does not independently re-annotate them or guarantee their correctness.

For the methodology and dataset construction, refer to our HateSieve paper. The source artifacts do not specify a new dataset license. No additional license is granted by this card; underlying materials remain subject to their respective source terms.

Citation

If you use this dataset in your research, please cite the accompanying paper:

@inproceedings{su-etal-2025-context,
    title = {A Context-Aware Contrastive Learning Framework for Hateful Meme Detection and Segmentation},
    author = {Su, Xuanyu and Li, Yansong and Inkpen, Diana and Japkowicz, Nathalie},
    booktitle = {Findings of the Association for Computational Linguistics: NAACL 2025},
    month = apr,
    year = {2025},
    address = {Albuquerque, New Mexico},
    publisher = {Association for Computational Linguistics},
    url = {https://aclanthology.org/2025.findings-naacl.289/},
    doi = {10.18653/v1/2025.findings-naacl.289},
    pages = {5216--5230}
}

Content and limitations

This is research data about hateful and offensive multimodal content. Text and images may contain offensive language, stereotypes, or other disturbing material. Labels reflect the source annotations and construction procedure, rather than endorsement of the content. Generated and retrieved examples can introduce annotation noise and artifacts.

Export validation

All 22,095 referenced image files were fully decoded and checked for RGB conversion. Every exported row was read back from Parquet and compared with its source: non-image fields match exactly and all image bytes match SHA-256 digests of the source files. The source cache-*.arrow preprocessing files are excluded.